The chapter's decision section. Sections 7.1 and 7.2 supplied the distributions of a sample mean and of a sum; this one supplies the question that must be answered before either can be used, namely whether the problem concerns an individual value, a mean, or a total. The book is explicit that a question about an individual value must be answered from that variable's own distribution and not from the central limit theorem at all. The section also states the law of large numbers, which says that as samples get larger their means get closer to the population mean, and observes that the central limit theorem illustrates it because the standard error shrinks toward zero. Two worked examples take populations that are conspicuously not normal, a uniform and an exponential, which is where the theorem does real work rather than restating what chapter 6 already covered.
Subject: Statistics · 69 slides · symbolic lesson
Open the interactive version of this deck
Title
Statistics · §7.3
An individual, a mean, or a total: choosing the right distribution, and what it means when you do
Objectives
Sections 7.1 and 7.2 built two distributions out of one population: one for a sample mean, one for a sample sum. This section is about using them, and above all about knowing which one a question needs, or whether it needs neither.
The arithmetic here is the same z-score and area work you already know from Chapter 6. What is new, and where the marks go, is the decision before the arithmetic and the sentence after it.
Warm-up
Everything in this section rests on four numbers you can get from any population. Recall them before going on, with a population whose mean is 50 and standard deviation 10, and samples of 25.
Discussion prompt
A population has mean 50 and standard deviation 10, and you take samples of size 25. What are the mean and standard deviation of the sample mean? What are the mean and standard deviation of the sample sum?
Keep one contrast in mind: dividing by the square root of n makes the mean's distribution narrower than the population, while multiplying makes the sum's wider. Mixing the two up is the commonest slip in this chapter, and a slide in Part 3 is built around exactly that error.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.1-7.2, recalled in §7.3, p. 374 — the standard deviation of the sample mean is σ over √n
Section
Part 1
Concept
Take the stress scores of Example 7.8: every student's score is equally likely to be anywhere from 1 to 5. That single population gives you three different distributions, depending on what you are asked about.
Figure (svg): Three panels. Left: a flat orange rectangle from 1 to 5, the uniform distribution of one stress score, labelled mu 3 and sigma 1.15. Middle: a tall narrow bell curve centred at 3 on an axis from 2.5 to 3.5, the distribution of the mean of 75 scores, N of 3 and 0.133. Right: a bell curve centred at 225 on an axis from 190 to 260, the distribution of the total of 75 scores, N of 225 and 10.
On the left is one student's score, flat from 1 to 5. In the middle is the mean score of 75 students, a tall narrow bell squeezed around 3. On the right is the total of those 75 scores, a bell around 225 with a spread of 10.
\[ X \sim U(1,5), \quad \bar{X} \sim N\left(3,\ \tfrac{1.15}{\sqrt{75}}\right), \quad \Sigma X \sim N\left(75(3),\ \sqrt{75}(1.15)\right) \]
Look at the three axes. The same question, is it below 2, has three very different answers: for one student it happens a quarter of the time, for a mean of 75 it essentially never happens, and for a total it is not even the right scale. So before you compute anything, you have to know which of the three the question is about.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 374 — Example 7.8 set-up
Notation
The two central limit theorem results side by side. Each piece is worth reading slowly, because the two lines look alike and do opposite things to the spread.
Annotate
On: \( \bar{X} \sim N\left(\mu,\ \frac{\sigma}{\sqrt{n}}\right) \qquad \Sigma X \sim N\left(n\mu,\ \sqrt{n}\,\sigma\right) \)
Notice that the total's standard deviation is exactly n times the mean's: multiply the standard error by n and you get root n times sigma. That is because the total is just n times the mean, so every number about a total is n times the matching number about a mean.
\[ n \cdot \frac{\sigma}{\sqrt{n}} = \sqrt{n}\,\sigma \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, pp. 373-374 — use the clt for the mean, or the clt for sums
Concept
The book puts the whole decision in three sentences. If you are asked about the probability of a mean, use the central limit theorem for the mean. If you are asked about a sum or total, use it for sums. The same goes for percentiles.
Figure (svg): A decision diagram. A box on the left reads: read the question, what is it about? Three arrows lead right. The orange one, one value, cued by the words one or a randomly chosen, leads to X's own distribution such as uniform, exponential or normal. The blue one, a mean, cued by mean or average, leads to x-bar distributed normal with mean mu and standard deviation sigma over root n. The yellow one, a sum, cued by total or sum, leads to the sum distributed normal with mean n mu and standard deviation root n times sigma.
And then a note in bold: if you are asked about an individual value, do not use the central limit theorem at all. Use the distribution of that value's own random variable, whatever it is, uniform, exponential or normal.
So read for the deciding word. Mean or average sends you to the middle branch. Total or sum sends you to the bottom branch. One, a single, or one randomly selected person sends you to the top, and there the population's own distribution is the answer, skewed or flat or whatever it happens to be.
Why is the individual branch off limits? Because the theorem is a statement about averages and totals of many values. A single value is not averaged with anything, so nothing smooths its distribution into a bell. The customer who goes over her minutes is one draw from an exponential, and she stays exponential.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, pp. 373-374 — when to use the clt, and the note on individual values
Concept
The mean and the total of the same sample are tied together: the total is always n times the mean. So every event about a total can be rewritten as an event about the mean, just by dividing the threshold by n.
\[ \Sigma X < 200 \iff \frac{\Sigma X}{75} < \frac{200}{75} \iff \bar{X} < 2.667 \]
Those three events are the same event, so they have the same probability. That is why the book can offer two theorems, one for means and one for sums, without them ever disagreeing. They are one theorem read on two rulers.
\[ z = \frac{200 - 225}{10} = -2.5 \qquad z = \frac{2.667 - 3}{0.1333} = -2.5 \]
Both routes give the same z-score, so it does not matter which one you use, as long as you keep the centre and spread that belong to the ruler you chose. What goes wrong is mixing them: a total's threshold with a mean's spread, or the other way round.
An individual question has no such translation. One student's score is not the total of anything, and dividing it by 75 does not turn it into a mean. That is the reason the individual branch sits apart from the other two in the diagram.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 373 — use the clt for the mean, or the clt for sums
Sorting
Each question below is about the stress scores, the excess cell phone minutes, or the Broadway ages. Sort it by the distribution you would use. Find the deciding word first.
Sort into buckets
Sort each question by the distribution it needs.
The individual questions are the ones people get wrong, because they look like every other question in the chapter. A single stress score above 4 has probability one quarter straight from the flat distribution: the part from 4 to 5 is a quarter of the width from 1 to 5.
\[ P(X > 4) = \frac{5 - 4}{5 - 1} = 0.25 \]
If you put the single stress score with the means, you probably reached for the theorem because the example it comes from uses it. The example does, for its means and totals; for one student, the flat distribution answers directly.
Percentile questions follow exactly the same rule. The 90th percentile of the total is a sum question, even though the word probability never appears.
Trap
Example 7.9: excess minutes are exponential with mean 22, so the standard deviation is also 22. A first attempt at the chance that one customer goes more than 20 minutes over reaches for a normal curve:
\[ z = \frac{20 - 22}{22} \approx -0.09 \]
\[ P(Z > -0.09) \approx 0.536 \]
Wrong. One customer is not a mean, and the population is not normal.
An individual comes from the population's own distribution. Here that is exponential, whose right-tail area has a formula of its own.
\[ P(X > 20) = e^{-20/22} \]
\[ = e^{-0.9091} \approx 0.4029 \]
The normal answer is off by more than 0.13, because the exponential piles most of its probability near zero and the normal curve does not. The book's rule protects you: for an individual value, use the distribution of its random variable, never the central limit theorem.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 377 — Example 7.9(b)
Section
Part 2
Concept
The book states the law of large numbers in one sentence: if you take samples of larger and larger size from any population, the mean of the sample tends to get closer and closer to the population mean μ.
Figure (svg): Three wiggly lines, in blue, orange and purple, each tracing the running mean of simulated stress scores as more and more scores are averaged, from 1 to 1000. They start far apart, between about 1.5 and 4.5, swing widely over the first fifty or so scores, and then all settle into a narrow band hugging a dashed yellow line at 3. After 1000 scores the three means are 2.940, 3.005, 3.006.
Each line is one simulated student survey. The first score might be 1.4 or 4.8, so the running mean starts anywhere. By a hundred scores every line is within a few tenths of 3, and by a thousand within a few hundredths. The population mean of the stress scores is 3, and that is where every line is headed.
law of large numbers — As the sample size n grows, the sample mean tends to get closer and closer to the population mean μ. The population mean is the value the sample means approach.
Notice what the law does not say. It does not promise that any particular sample is close; a single unlucky survey can still wander. It says that the bigger the sample, the less room there is to wander.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 374 — Law of Large Numbers
Concept
The book's argument is short. The central limit theorem says sample means are approximately normal, centred at μ, with standard deviation sigma over root n. Watch what happens to that spread as n grows, for the stress scores, whose sigma is 1.1547.
\[ \frac{1.1547}{\sqrt{1}} = 1.155, \quad \frac{1.1547}{\sqrt{25}} = 0.231, \quad \frac{1.1547}{\sqrt{400}} = 0.058 \]
The spread shrinks toward zero. A normal curve with almost no spread has almost all of its area right next to its centre, and its centre is μ. So the sample mean must be close to μ: that is the law of large numbers, and the theorem has delivered it.
Figure (svg): Two curves open to the left and close in to the right like a funnel around a dashed line at 3: the upper curve is 3 plus two standard errors, the lower is 3 minus two standard errors, for n from 1 to 400. Forty white dots, one simulated sample mean for each n from 10 to 400 in steps of 10, sit inside the funnel and crowd closer to 3 as n grows.
The funnel is two standard errors either side of 3. Each white dot is a real simulated sample mean, and they obey it: scattered when n is small, squeezed onto the line when n is large. Quadrupling the sample halves the funnel's width, because the square root of four is two.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 374 — the central limit theorem illustrates the law of large numbers
Intuition
It is tempting to think that averaging n values should shrink the spread by a factor of n. It shrinks it only by the square root of n, and the reason is that random errors partly cancel rather than all pulling the same way.
Work with variances, which add for independent values. A total of n independent scores has n times the variance of one score.
\[ \text{Var}(\Sigma X) = n\sigma^{2} \]
Take the square root to get back to a standard deviation. The n comes out as root n.
\[ \sigma_{\Sigma x} = \sqrt{n\sigma^{2}} = \sqrt{n}\,\sigma \]
The mean is the total divided by n, so its spread is the total's spread divided by n.
\[ \sigma_{\bar{x}} = \frac{\sqrt{n}\,\sigma}{n} = \frac{\sigma}{\sqrt{n}} \]
So the total's spread grows, but more slowly than the total itself, because some students score high and some low and those surpluses partly cancel. Dividing by n then more than undoes that growth for the mean. The price is steep: to halve the standard error you need four times the sample, and to cut it to a tenth you need a hundred times the sample.
Concept
Make the law concrete with one question: how likely is a sample mean of stress scores to land within 0.1 of the true mean, 3? The answer is the area of the bell between 2.9 and 3.1, and it depends only on the standard error.
\[ P(2.9 < \bar{X} < 3.1) = P\left(|Z| < \frac{0.1}{\sigma/\sqrt{n}}\right) \]
| sample size n | standard error σ/√n | 0.1 in standard errors | P(within 0.1 of 3) |
|---|---|---|---|
| 25 | 0.231 | 0.43 | 0.335 |
| 100 | 0.115 | 0.87 | 0.614 |
| 400 | 0.058 | 1.73 | 0.917 |
| 1600 | 0.029 | 3.46 | 0.9995 |
Read down the last column. With 25 students the survey lands within a tenth of the truth only a third of the time. With 1600 it almost always does. Each row quadruples n, which doubles the number of standard errors that 0.1 is worth, and the probability climbs toward one.
That is the law of large numbers in numbers rather than words. It is also the reason national polls survey a thousand or more people and not twenty-five: the extra people buy closeness, at a rate of the square root.
Tweak it
The slider draws the distribution of the mean of n stress scores: a normal curve centred at 3 with standard error 1.1547 over root n. Move n and watch where the area goes.
Parameter explorer
Start at n equal to 4 and read roughly how wide the curve is at its base. Then move to 16, 64 and 256. Each time n is multiplied by four, what happens to the width and to the height of the curve?
\[ \bar{X} \sim N\left(3,\ \frac{1.1547}{\sqrt{{n}}}\right) \]
Watch the height as well as the width. The area under the curve is always one, so as the curve narrows it must grow taller; at n equal to 100 the peak is about 3.5, far above anything a probability could be. That is allowed, because a density height is not a probability.
\[ n \times 4 \Rightarrow \frac{\sigma}{\sqrt{4n}} = \frac{1}{2}\cdot\frac{\sigma}{\sqrt{n}} \]
Notice that the peak never leaves 3, whatever n you choose: the centre of the sample mean's distribution is always μ. Only the spread depends on n, which is why a bigger sample makes the mean more trustworthy but never shifts what it is aiming at.
Counterexample
A common reading of the law of large numbers goes like this: a fair coin has come up heads ten times in a row, so tails is now due, because in the long run the proportion must come back to one half. Test that claim against the picture.
Figure (svg): A curve of the expected proportion of heads after an opening run of 10 straight heads followed by fair flips, plotted against the total number of flips from 10 to 1000. It starts at 1, drops steeply, and flattens toward a dashed line at 0.5, reaching 0.505 at 1000 flips.
Discussion prompt
After 10 straight heads, a fair coin is flipped 990 more times. What is the expected proportion of heads over all 1000 flips? Does the coin need to produce extra tails for the proportion to approach one half?
The picture makes the point without any algebra. The curve falls toward one half steadily and never dips below it, which is what would have to happen if extra tails were being produced to make up for the early heads.
So the law of large numbers is not a force that corrects past results. Each flip is independent and has no memory. The proportion approaches one half because the early surplus is swamped by the growing number of fair flips, exactly as the running means in the first picture of this part were swamped.
Prediction
Two students each survey stress scores. One averages 10 scores, the other 100. Before computing, predict which of them more often gets a mean within 0.2 of the true 3, and by roughly how much.
Predict first
Which sample mean lands within 0.2 of 3 more often?
Correct: the mean of 100: about 92 percent of the time, against about 42 percent for the mean of 10
Why: Both distributions are centred at 3, but the standard error for 100 scores is 0.115 and for 10 scores it is 0.365. So 0.2 is about 1.73 standard errors for the bigger sample and only 0.55 for the smaller one, and the probabilities are about 0.92 and 0.42.
\[ \frac{0.2}{1.1547/\sqrt{10}} = 0.55, \qquad \frac{0.2}{1.1547/\sqrt{100}} = 1.73 \]
The option that says both are equally likely is the trap: both curves are centred at 3, but having the right centre is not the same as being close. Closeness is about spread, and spread is what the sample size buys.
Section
Part 3
Concept
Example 7.8 studies stress on a college campus. Stress scores follow a uniform distribution from a lowest score of 1 to a highest of 5. Let X be one student's stress score. Before any central limit theorem, you need the population's own mean and standard deviation, from the uniform formulas of Chapter 5.
\[ X \sim U(1, 5), \quad a = 1, \ b = 5 \]
\[ \mu = \frac{a + b}{2} = \frac{1 + 5}{2} = 3 \]
The mean is the midpoint, as symmetry demands. The standard deviation comes from the width of the interval, squared, over twelve.
\[ \sigma = \sqrt{\frac{(b-a)^{2}}{12}} = \sqrt{\frac{16}{12}} = \sqrt{\frac{4}{3}} \approx 1.1547 \]
The book rounds this to 1.15. The deck keeps 1.1547, because then the numbers come out exact: the standard error for 75 students is exactly two fifteenths and the total's standard deviation is exactly 10. The answers agree with the book's to the places it prints.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 374 — Example 7.8: μ = 3, σ ≈ 1.15
Picture it
This is where the theorem earns its keep. Nothing about the stress scores is bell-shaped: every value from 1 to 5 is equally likely. Yet here are 2000 simulated surveys of 75 students, each reduced to its mean.
Figure (svg): Left: the flat uniform density of one stress score, a rectangle from 1 to 5. An arrow labelled n equals 75 points right to a histogram of 2000 simulated means of 75 scores, piled in a bell shape between about 2.6 and 3.4, with a yellow normal curve N of 3 and 0.133 drawn over the bars. The 2000 means average 2.998 with standard deviation 0.131.
The means pile into a bell between about 2.6 and 3.4, and the yellow curve, drawn from the theorem alone before any simulation, fits the bars. The 2000 simulated means averaged 2.998, with standard deviation 0.131, against the theorem's 3 and 0.133.
Two things changed between the panels. The shape went from flat to bell, which is the normal part of the theorem. And the spread collapsed from the whole range 1 to 5 down to a band less than one unit wide, which is the square root of 75 at work.
Keep this picture in mind for part (a) of the example. A mean of 2 is not on the right-hand panel at all.
Worked example
Using a sample of 75 students, parts (a) and (b) ask about the mean stress score, and parts (c) and (d) about the total. Set up both distributions first, one move per line.
\[ n = 75, \quad \mu = 3, \quad \sigma = \sqrt{4/3} \approx 1.1547 \]
Centre of the mean
Why: a sample mean is centred at the population mean.
\[ \mu_{\bar{x}} = 3 \]
The mean first. Its centre does not move: averaging many scores from a population centred at 3 gives averages centred at 3. Only the spread changes, and it changes by the square root of the sample size, not by the sample size itself.
Standard error of the mean
Why: divide σ by the square root of n.
\[ \sigma_{\bar{x}} = \frac{1.1547}{\sqrt{75}} = \frac{1.1547}{8.660} \]
Simplify
Why: exactly two fifteenths.
\[ \sigma_{\bar{x}} = 0.1333 \]
So the mean stress score of 75 students is approximately normal with mean 3 and standard error 0.1333. Now the total: the centre is multiplied by n and the spread by the square root of n.
Centre of the total
Why: n times μ.
\[ \mu_{\Sigma x} = 75(3) = 225 \]
Spread of the total
Why: root n times σ; the book writes it as root 75 times root four thirds.
\[ \sigma_{\Sigma x} = \sqrt{75}\sqrt{\tfrac{4}{3}} = \sqrt{100} = 10 \]
\[ \bar{X} \sim N(3,\ 0.1333), \qquad \Sigma X \sim N(225,\ 10) \]
Is n equal to 75 large enough? Comfortably: the uniform is symmetric with no long tail, so even samples of a dozen give nearly normal means. The picture on the previous slide shows it working at 75.
Check that the total's spread is n times the mean's
Why: the total is n times the mean, so this must hold exactly.
\[ 75 \times 0.1333 = 10.0 = \sigma_{\Sigma x} \quad \checkmark \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, pp. 374-375 — Example 7.8 set-up
Worked example
Find the probability that the mean stress score for the 75 students is less than 2. This is a question about a mean, so it uses the normal curve with mean 3 and standard error 0.1333.
\[ P(\bar{X} < 2), \quad \bar{X} \sim N(3,\ 0.1333) \]
The threshold 2 looks close to 3 on the stress scale, just one point away. What matters is how many standard errors that point is worth, and with a standard error of 0.1333 a single point is a very long way.
Standardise
Why: distance from 3, in standard errors.
\[ z = \frac{2 - 3}{0.1333} \]
Divide
Why: a whole unit is seven and a half standard errors.
\[ z = \frac{-1}{0.1333} = -7.5 \]
A z-score of minus 7.5 is off every table in the back of the book, which stops at about minus 3.5. That is itself the answer: anything beyond about four standard errors has an area too small to matter.
Read the area
Why: left of z equal to minus 7.5 on the standard normal.
\[ P(Z < -7.5) \approx 3 \times 10^{-14} \approx 0 \]
Figure (svg): A narrow bell curve for the mean stress score, centred at 3 on an axis from 1.8 to 3.6, with a z row underneath reading minus 7.5 at 2, minus 3 at 2.6, 0 at 3 and 3 at 3.4. A red line at 2 stands where the curve has long since reached the axis; a red arrow from near 3 back to 2 is labelled 7.5 standard errors.
So the probability that the mean stress score of 75 students is below 2 is essentially zero. The book's calculator line starts the area at 1, normalcdf(1, 2, 3, 1.15 over root 75), as a reminder that no score can be below 1, but the answer is zero either way.
Why so extreme, when a single student scores below 2 a quarter of the time? For the mean to be below 2, the 75 scores would all have to lean low together, and with independent students that essentially never happens. The book's own Figure 7.4 draws 2 just under the left shoulder of the curve, which undersells this badly: at 7.5 standard errors out, the curve has been flat on the axis for a long way.
Check with the individual
Why: one score below 2 is a quarter of the flat distribution, so the mean's near-zero answer is a real contrast, not an arithmetic slip.
\[ P(X < 2) = \frac{2-1}{4} = 0.25 \quad \text{vs} \quad P(\bar{X} < 2) \approx 0 \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, pp. 374-375 — Example 7.8(a), Figure 7.4
Worked example
Find the 90th percentile for the mean of 75 stress scores. A percentile runs the z-score backwards: you know the area, 0.90 to the left, and want the cut k.
\[ P(\bar{X} < k) = 0.90 \]
Work in two stages: first find where the cut is on the standard normal curve, then translate it to the mean's scale. The first stage does not care about stress scores at all; 1.2816 is the 90th percentile of every normal curve, counted in standard deviations.
Find the z with 0.90 to its left
Why: from the standard normal table, or invNorm(0.90, 0, 1).
\[ z_{0.90} = 1.2816 \]
Convert back to the mean's scale
Why: start at 3 and walk z standard errors.
\[ k = 3 + (1.2816)(0.1333) \]
Multiply
Why: 1.28 standard errors, in stress-score units.
\[ k = 3 + 0.1709 \]
Add
Why: the book rounds to one decimal place.
\[ k = 3.1709 \approx 3.2 \]
Figure (svg): The bell curve of the mean stress score, centred at 3, with the area to the left of 3.17 shaded blue and labelled area 0.90; the small unshaded right tail is labelled 0.10. The z row shows 3.17 sits at z about 1.28.
Interpret it in context, the way the book does: 90 percent of all the means of 75 stress scores are at most 3.2, and 10 percent are at least 3.2. It is a statement about means of samples of 75, not about individual students.
Compare it with one student: the 90th percentile of a single score is 1 plus 0.9 of the way to 5, which is 4.6. The mean's cut is barely above the centre because the mean's spread is so small.
Check by standardising the answer
Why: the cut should sit 1.28 standard errors above 3.
\[ \frac{3.1709 - 3}{0.1333} = 1.28 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 375 — Example 7.8(b), Figure 7.5
Worked example
Find the probability that the total of the 75 stress scores is less than 200. The word total sends you to the sum's distribution, normal with mean 225 and standard deviation 10.
\[ P(\Sigma X < 200), \quad \Sigma X \sim N(225,\ 10) \]
Before computing, place 200 on the picture in your head: the total is centred at 225 with a spread of 10, so 200 is a couple of spreads below. Expect a small probability, but not a vanishing one like part (a).
Standardise
Why: distance from 225, in standard deviations of the total.
\[ z = \frac{200 - 225}{10} \]
Divide
Why: 25 below is two and a half standard deviations.
\[ z = \frac{-25}{10} = -2.5 \]
Read the area
Why: left of z equal to minus 2.5; the calculator form is normalcdf(75, 200, 225, 10).
\[ P(Z < -2.5) = 0.0062 \]
Figure (svg): A bell curve for the total of 75 stress scores, centred at 225 with standard deviation 10, on an axis from 185 to 265, with a z row: 200 sits at z minus 2.5. The thin left tail below 200 is shaded red and labelled area 0.0062.
So the probability that the total of 75 stress scores is below 200 is about 0.0062, roughly six chances in a thousand. The lower limit 75 in the calculator line is the smallest possible total, 75 students each scoring 1.
Two misprints in the book here. The solution line prints the probability as minus 0.0062, and a probability can never be negative; the value is 0.0062, as the next sentence of the book says. And Figure 7.6 is labelled with a probability of 0, which contradicts that same sentence.
Check on the mean's scale
Why: a total of 200 is a mean of 200 over 75; it must give the same z.
\[ \frac{200}{75} = 2.667, \quad \frac{2.667 - 3}{0.1333} = -2.5 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, pp. 375-376 — Example 7.8(c), Figure 7.6
Worked example
Find the 90th percentile for the total of the 75 stress scores. Same z as part (b), because it is the same 0.90 of area; only the scale changes.
\[ P(\Sigma X < k) = 0.90, \quad \Sigma X \sim N(225,\ 10) \]
Use the z with 0.90 to its left
Why: as in part (b).
\[ z_{0.90} = 1.2816 \]
The z-value is the same as in part (b) because both parts ask for the same 90 percent of area. What differs is the ruler it is laid against: standard errors of 0.1333 for the mean, standard deviations of 10 for the total.
Convert to the total's scale
Why: start at 225 and walk z standard deviations of 10.
\[ k = 225 + (1.2816)(10) \]
Multiply and add
Why: the calculator form is invNorm(0.90, 225, 10).
\[ k = 225 + 12.816 = 237.8 \]
Figure (svg): The bell curve of the total of 75 stress scores with the area left of 237.8 shaded and labelled 0.90. Under the axis ticks 195, 210, 225, 237.8 and 255 a second row divides each by 75, reading 2.6, 2.8, 3, 3.17 and 3.4.
In context: 90 percent of all totals of 75 stress scores are no more than 237.8, and 10 percent are no less. If you use the book's rounded sigma of 1.15, the total's spread is 9.96 instead of 10 and the cut is 237.76, which still rounds to 237.8.
The picture shows why parts (b) and (d) are one question on two rulers. Dividing the total's axis by 75 turns it into the mean's axis, cut for cut.
Check against part (b)
Why: the total's percentile divided by n must be the mean's percentile.
\[ \frac{237.8}{75} = 3.17 = k_{\bar{x}} \quad \checkmark \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 376 — Example 7.8(d), Figure 7.7
Error analysis
Four students set up the distribution of the mean of 75 stress scores. Only one used the right spread. Find the fault in each of the others; every one of these slips is common.
Annotate
On: \( \frac{1.15}{75} \quad \frac{1.15^{2}}{75} \quad \sqrt{75}(1.15) \quad \frac{1.15}{\sqrt{75}} \)
The first two errors make the mean look far more precise than it is. With 0.0153 as the spread, a mean of 3.02 would already be more than one standard error from 3, which is nonsense for a survey of 75 people.
A quick way to catch all three: the mean of 75 scores should be within a few tenths of 3 almost every time, since the scores run from 1 to 5. A standard error of 0.0153 says within a few hundredths, and 9.96 says anywhere at all; only 0.133 matches that sense of scale.
The third error is the one this section invites, because parts (a) to (d) use both spreads side by side. Keep them attached to their centres: 3 goes with 0.133, and 225 goes with 10. If a spread and a centre are on different scales, one of them is wrong.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, pp. 374-375 — Example 7.8
Anomaly
Try It 7.8 keeps the stress scores but uses 55 students. Its first part asks for the probability that the mean stress score is less than 7. Before reaching for a z-score, think about what the scores can be.
Predict first
What is the probability that the mean of 55 stress scores is less than 7?
Correct: 1
Why: Every stress score is at most 5, so every mean of stress scores is at most 5, and 5 is less than 7. The event is certain. The normal curve agrees: 7 is about 26 standard errors above 3.
\[ \sigma_{\bar{x}} = \frac{1.1547}{\sqrt{55}} = 0.1557, \quad z = \frac{7 - 3}{0.1557} \approx 25.7 \]
A question like this is checking whether you look at the scale before computing. The mean of values between 1 and 5 is always between 1 and 5, so a threshold outside that range decides the answer on its own: probability 1 here, and 0 for a question like the mean exceeding 7.
The option about 0 comes from misreading the direction, and the option saying it cannot be found forgets that you do not need the curve at all when the event is certain.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 376 — Try It 7.8(a)
Fill the middle
With 55 students the total has mean 165 and standard deviation root 55 times 1.1547, which is 8.563; the mean has standard error 0.1557. Fill in the three results, rounding as shown. The z-values are 0.5839 for part (b), 0.8416 for the 80th percentile and 1.0364 for the 85th.
Fill in the blanks
P(\Sigma X > 170) = P(Z > 0.58) = 0.2797 \qquad k_3.13} = 3 + 0.8416(0.1557) = 173.88 \qquad k____ = 165 + 1.0364(8.563) = ___
Why: Part b uses the total's curve: 170 is 5 above 165, which is 0.58 of a standard deviation of 8.563, and the area to its right is 0.2797. Part c walks 0.8416 standard errors above 3 on the mean's scale. Part d walks 1.0364 standard deviations above 165 on the total's scale.
Each blank uses the curve its wording demands: the sum for parts (b) and (d), the mean for part (c). A total above 170 is not rare at all here, happening more than a quarter of the time, because 170 is barely half a standard deviation above the centre of 165.
\[ \frac{173.88}{55} = 3.16 \]
The line above is a quick sanity check on part (d): the 85th percentile of the total, divided by 55, lands a little above the 80th percentile of the mean, 3.13, as it should for a higher percentile.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 376 — Try It 7.8(b)-(d)
Section
Part 4
Picture it
Example 7.9 is about cell phone customers who go over the minutes in their basic contract. For those customers, the excess time is exponential with mean 22 minutes. An exponential's standard deviation equals its mean, so sigma is 22 as well.
\[ X \sim \text{Exp}\left(\tfrac{1}{22}\right), \quad \mu = 22, \quad \sigma = 22 \]
Figure (svg): Two densities on one axis of excess minutes from 0 to 80. An orange exponential curve starts at its highest point at zero and decays slowly to the right, with everything right of 20 shaded orange and labelled P of X greater than 20 equals 0.40. A tall narrow blue bell centred at 22 has its area right of 20 shaded blue, labelled P of x-bar greater than 20 equals 0.79. A white line marks 20.
The orange curve is one customer: most customers go only a little over, but a few go a very long way, and that long right tail is what skewed means. The blue bell is the mean of 80 such customers: centred at 22, with a standard error of only 2.46 minutes.
Now look at the threshold of 20 minutes. More than half of the orange area lies left of 20, but most of the blue area lies right of it. So the answer to is it more than 20 depends entirely on whether the question is about one customer or about the average of 80, and the next slide computes both.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, pp. 376-377 — Example 7.9
Concept
Questions about one customer need the exponential itself, so recall how it works. With mean 22 minutes, the decay rate is m, one over the mean, and the chance of exceeding any number of minutes x has a one-line formula.
\[ m = \frac{1}{\mu} = \frac{1}{22} = 0.04545 \]
\[ P(X > x) = e^{-mx} = e^{-x/22} \]
That formula gives more than probabilities. Setting it equal to one half finds the median, the number of minutes that half of customers exceed.
\[ e^{-x/22} = 0.5 \Rightarrow x = 22\ln 2 = 15.25 \text{ minutes} \]
The median, 15.25, is well below the mean, 22. That gap is what a right skew looks like in numbers: most customers go over by a modest amount, and a few go over by a great deal, dragging the mean up. For a normal curve the median and the mean would be equal.
Keep 15.25 in mind. The threshold in Example 7.9 is 20 minutes, which is above the median of one customer but below the centre of the mean's bell. That single fact predicts which of the two probabilities will be bigger before you compute either.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 376 — Example 7.9: X ~ Exp(1/22), μ = σ = 22
Worked example
(a) Find the probability that the mean excess time of a sample of 80 customers is more than 20 minutes. (b) Find the probability that one randomly selected customer's excess time is more than 20 minutes.
Part (a) is about a mean of 80, a large sample, so the theorem applies even though the population is strongly skewed. The centre stays at 22 minutes and the spread shrinks from 22 to 22 over root 80.
(a) Standard error of the mean
Why: sigma over root 80.
\[ \sigma_{\bar{x}} = \frac{22}{\sqrt{80}} = \frac{22}{8.944} = 2.4597 \]
Standardise 20
Why: distance from 22, in standard errors.
\[ z = \frac{20 - 22}{2.4597} = -0.813 \]
Area to the right
Why: the calculator form is normalcdf(20, 1E99, 22, 22 over root 80).
\[ P(\bar{X} > 20) = P(Z > -0.813) = 0.7919 \]
In context: there is a 0.7919 probability that the mean excess time of 80 such customers is more than 20 minutes. The book prints 0.79199 and then 0.7919; node gives 0.79192, so 0.7919 is right. Now part (b), which is about one customer, so the central limit theorem does not apply.
(b) Use the exponential's tail
Why: for an exponential, the chance of exceeding x is e to the minus x over the mean.
\[ P(X > 20) = e^{-20/22} \]
Simplify the exponent
Why: the book writes it with the decay rate 1/22 = 0.04545.
\[ e^{-(0.04545)(20)} = e^{-0.909} \]
Evaluate
Why: about two customers in five go more than 20 minutes over.
\[ P(X > 20) = 0.4029 \]
Part (c) asks why the answers differ. They use different distributions: one customer is a single draw from a skewed exponential whose median is only 15.2 minutes, while the mean of 80 comes from a narrow bell centred at 22.
Check the direction of each answer
Why: 20 is below 22, so the mean should exceed it more often than not; and for one customer, 20 is above the exponential's median, 22 times ln 2, so fewer than half should exceed it.
\[ 0.7919 > 0.5, \qquad 22\ln 2 = 15.25 < 20 \Rightarrow 0.4029 < 0.5 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 377 — Example 7.9(a)-(c)
Comparison
The book answers part (c) in three numbered lines. Fill in the table to make each line concrete: every row is one thing that differs between the two calculations.
Comparison matrix
| one customer, X | mean of 80, x̄ | |
|---|---|---|
| distribution | exponential | normal |
| centre | 22 | 22 |
| standard deviation | 22 | 2.46 |
| median | 15.25 | 22 |
| P(more than 20) | 0.4029 | 0.7919 |
The centres are the same, 22 minutes. Everything else differs: the shape, the spread, and so the median and the tail area. The mean of 80 is symmetric, so its median is its centre, 22, and 20 sits below it. One customer's distribution is skewed right, so its median of 15.25 is well below its mean, and 20 sits above it.
That is the whole of the book's point (c). The two probabilities are not supposed to be equal, because they answer different questions with different distributions. For an individual, use the stated distribution of its random variable; for a mean, use the central limit theorem.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 377 — Example 7.9(c)
Worked example
Find the 95th percentile of the sample mean excess time for samples of 80 customers. It is a percentile of a mean, so it uses the bell with centre 22 and standard error 2.4597.
\[ P(\bar{X} < k) = 0.95, \quad \bar{X} \sim N(22,\ 2.4597) \]
Find the z with 0.95 to its left
Why: from the table, or invNorm(0.95, 0, 1).
\[ z_{0.95} = 1.6449 \]
The 95th percentile uses the z-value 1.6449, which leaves exactly 5 percent above it. It is the same z you would use for any normal quantity; the context enters only when you convert back to minutes.
Convert to minutes
Why: start at 22 and walk z standard errors.
\[ k = 22 + (1.6449)(2.4597) \]
Multiply
Why: 1.64 standard errors is about four minutes.
\[ k = 22 + 4.046 \]
Add and round
Why: the book's answer.
\[ k = 26.05 \approx 26.0 \text{ minutes} \]
Figure (svg): The bell curve of the mean excess time for samples of 80, centred at 22 with standard error 2.46, on an axis from 14 to 30 with a z row. The right tail beyond 26.05, where z is 1.64, is shaded yellow and labelled top 5 percent of means; the rest of the area is labelled 0.95.
In context: 95 percent of samples of 80 such customers have a mean excess time under 26 minutes, and only 5 percent have a mean above it. It says nothing about individual customers; in fact about 31 percent of individual customers go more than 26 minutes over, since e to the minus 26 over 22 is 0.306.
Check by standardising the cut
Why: 26.05 should sit 1.645 standard errors above 22.
\[ \frac{26.05 - 22}{2.4597} = 1.646 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, pp. 377-378 — Example 7.9, Figure 7.9
Ranking
Stay with the excess minutes, mean 22 and sigma 22. Four questions ask how likely it is to exceed 26 minutes. Rank them from most likely to least likely, using what you know about how the spread depends on n.
Put in order
Rank from most likely to least likely.
Why: The threshold 26 is above the centre 22 in every case, so a smaller spread means a smaller chance. One customer has the widest distribution and a long right tail, about 0.31. The means have standard errors of 4.02, 2.46 and 1.83, so 26 is 1.0, 1.6 and 2.2 standard errors up, giving about 0.16, 0.052 and 0.015.
\[ \frac{22}{\sqrt{30}} = 4.02, \quad \frac{22}{\sqrt{80}} = 2.46, \quad \frac{22}{\sqrt{144}} = 1.83 \]
The pattern is the law of large numbers again. The bigger the sample, the less its mean can stray from 22, so the less often it reaches as far as 26. The individual comes first not because 26 is special for one customer, but because a single customer is not averaged with anyone.
If a threshold sat below 22 instead, say 18, the order would reverse: bigger samples would exceed it more often, because their means hug 22.
Worked example
Use Example 7.9 with samples of 144. Find (a) the probability the mean is between 20 and 30, (b) the probability the sum is at least 3000, (c) the 75th percentile of the mean, (d) the 85th percentile of the sum.
Spread of the mean
Why: 22 over 12.
\[ \sigma_{\bar{x}} = \frac{22}{\sqrt{144}} = \frac{22}{12} = 1.8333 \]
(a) Standardise both ends
Why: 20 and 30 on the mean's scale.
\[ z_{1} = \frac{20-22}{1.8333} = -1.09, \quad z_{2} = \frac{30-22}{1.8333} = 4.36 \]
Area between
Why: almost everything above minus 1.09.
\[ P(20 < \bar{X} < 30) = 0.8623 \]
Part (a) was a mean, so it used the standard error. Part (b) switches to a sum, and with it both the centre and the spread change scale: the centre is multiplied by 144, and the spread by 12, the square root of 144.
(b) Centre and spread of the sum
Why: n times 22, and 12 times 22.
\[ \mu_{\Sigma x} = 144(22) = 3168, \quad \sigma_{\Sigma x} = 12(22) = 264 \]
Standardise 3000
Why: at least 3000 is the area to the right.
\[ z = \frac{3000 - 3168}{264} = -0.636 \]
Area to the right
Why: more than half, since 3000 is below the centre.
\[ P(\Sigma X \ge 3000) = 0.7377 \]
Parts (c) and (d) run the z-score backwards. Keep track of which ruler each uses: part (c) is a mean, with spread 1.8333, and part (d) a sum, with spread 264.
(c) 75th percentile of the mean
Why: z of 0.6745.
\[ k = 22 + 0.6745(1.8333) = 23.24 \text{ min} \]
(d) 85th percentile of the sum
Why: z of 1.0364.
\[ k = 3168 + 1.0364(264) = 3441.6 \text{ min} \]
Interpreted: 75 percent of samples of 144 customers have a mean excess time of at most 23.24 minutes, and 85 percent have a total of at most 3441.6 minutes. With 144 customers, about 95 percent of sample means land between about 18.3 and 25.7 minutes, two standard errors either side of 22.
Check (b) on the mean's scale
Why: a total of 3000 is a mean of 3000 over 144, which must give the same z.
\[ \frac{3000}{144} = 20.83, \quad \frac{20.83 - 22}{1.8333} = -0.636 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 378 — Try It 7.9
Worked example
Systolic blood pressure for women aged 18 to 24 is normal with mean 114.8 and standard deviation 13.1 mm Hg. Find the probability it is above 120 for (a) one woman, (b) the mean of 40 women.
(a) One woman: her own distribution
Why: which here happens to be normal.
\[ z = \frac{120 - 114.8}{13.1} = 0.397 \]
Area to the right
Why: about a third of women.
\[ P(X > 120) = 0.3457 \]
Notice that part (a) is an individual question and still uses a normal curve. That is not the central limit theorem: it is the population's own distribution, which the problem states is normal. The rule was always to use the individual's own distribution, whatever it is.
Now the mean of 40. The population is normal, so the mean is exactly normal, for any n; the only change is the spread.
(b) Standard error
Why: 13.1 over root 40.
\[ \sigma_{\bar{x}} = \frac{13.1}{\sqrt{40}} = 2.0713 \]
The distance from 114.8 to 120 is the same 5.2 mm Hg as in part (a). All that changes is the ruler: 5.2 is less than half of 13.1 but more than twice 2.07.
Standardise 120
Why: now 5.2 is two and a half standard errors.
\[ z = \frac{120 - 114.8}{2.0713} = 2.51 \]
Area to the right
Why: about six in a thousand.
\[ P(\bar{X} > 120) = 0.0060 \]
Figure (svg): Two normal curves centred at 114.8 on an axis of systolic blood pressure from 70 to 160. The low wide orange curve for one woman has its area right of 120 shaded, labelled 0.3457. The tall narrow blue curve for the mean of 40 women has only a sliver right of 120 shaded, labelled 0.0060. A white line marks 120.
In context: about 35 percent of women in this group have a systolic pressure above 120, but a group of 40 randomly chosen women would have an average above 120 only about 0.6 percent of the time. A sample average above 120 would be real evidence that the group is different from this population; a single reading above 120 would not.
Check the ratio of the z-scores
Why: the mean's z must be root 40 times the individual's, because only the spread changed.
\[ 0.397 \times \sqrt{40} = 0.397 \times 6.325 = 2.51 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 379 — Try It 7.10(a)(b)
Concept
The central limit theorem is a statement about large samples, so every use of it carries a condition. The section's examples run through three different situations, and it helps to keep them apart.
| population | is the sample mean normal? | example in this section |
|---|---|---|
| normal | yes, exactly, for any n | blood pressures, Try It 7.10 |
| not normal, n large (about 30 or more) | approximately, by the theorem | stress scores n = 75, excess minutes n = 80 |
| unknown or skewed, n small | no: cannot assume it | Try It 7.10(c), n = 4 |
The first row needs no theorem at all: averages of normal values are always normal. The second row is where the theorem does its work, turning flat or skewed populations into bell-shaped means. The third row is where you must stop, because a mean of a few skewed values is still skewed.
\[ n \ge 30 \quad \text{(a rule of thumb, larger for very skewed populations)} \]
The number 30 is a convention, not a law. A symmetric population like the stress scores behaves well with far fewer; a strongly skewed one like the excess minutes needs about 30 before the skew fades. The next slide lets you watch that happen.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 379 — Try It 7.10
Edge cases
Try It 7.10 ends by asking: if the sample were only four women and the original distribution were unknown, could the central limit theorem be used? Push the question to its limit with the most skewed population in this section, the excess minutes.
Figure (svg): Three exact densities for the mean of n exponential excess times with mean 22. For n equal to 1 the orange curve is the exponential itself, highest at zero. For n equal to 4 the purple curve has a hump near 16 and a long right tail; a dashed purple normal curve drawn for comparison is symmetric and spills below zero. For n equal to 30 the blue curve is a nearly symmetric bell centred near 22.
Parameter explorer
The curve is the exact distribution of the mean of n exponential excess times with mean 22. Start at n equal to 1 and increase it. At about what n does the hump look symmetric? What does the curve look like at n equal to 4?
\[ \bar{X} \text{ for } n = {n} \text{ exponential values, mean } 22 \]
So the answer to Try It 7.10(c) is no. With an unknown population and only four values, you cannot assume the mean is normal: if the population is skewed like this one, the mean of four is skewed too. The usual rule of thumb is n of at least 30, and more for very skewed populations. For the blood pressures it did not matter, because that population was stated to be normal, and the mean of a normal population is normal for any n.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 379 — Try It 7.10(c)
Real world
A Boeing 757 carries 200 passengers and has doors 72 inches high. Suppose men's heights are normal with mean 69.0 and standard deviation 2.8 inches, and half the passengers are men. The book asks two design questions and then which one an engineer should use.
Figure (svg): Two panels on the same scale of height in inches, 62 to 78. Top: the wide orange curve for one man's height, N of 69 and 2.8, shaded up to 73.6, where 95 percent of men fit. Bottom: the very narrow blue curve for the mean height of 100 men, N of 69 and 0.28, shaded up to 69.46. A dashed red line at 72 inches, the real door, crosses both panels.
Discussion prompt
(a) What doorway height lets 95 percent of men through without bending? (b) What height has a 0.95 probability of being above the mean height of 100 men? (c) Which result matters to the engineers designing the door, and why?
Read the figure's two panels together. The wide top curve is how tall the men who use the door actually are. The narrow bottom curve is how tall the average of a planeload of them is, which is useful for estimating a total but useless for the height of a door.
This is the book's rule in its most practical form. A design has to fit individuals, so it is an individual question, and the central limit theorem, which describes averages, would give a dangerously optimistic answer. Averages are for questions about groups, such as the total weight the aircraft carries.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 379 — Try It 7.11
Section
Part 5
Concept
Example 7.8 found the 90th percentile of the mean, 3.17, and of the total, 237.8, and the second was 75 times the first. That is not a coincidence, and you can see why in three lines.
\[ k_{\Sigma x} = n\mu + z\sqrt{n}\,\sigma \]
Factor n out of both terms. The second term needs root n times sigma written as n times sigma over root n.
\[ k_{\Sigma x} = n\left(\mu + z\frac{\sigma}{\sqrt{n}}\right) \]
\[ k_{\Sigma x} = n \cdot k_{\bar{x}} \]
So any percentile of a total is n times the same percentile of the mean, and the same goes for quartiles, the median and the interquartile range. This gives you a free check on every sum problem: divide by n and see whether you get a sensible mean.
It also tells you what scale each answer lives on. A mean of stress scores must be between 1 and 5; a total of 75 of them must be between 75 and 375. An answer outside its range means you used the wrong distribution.
Notation
Percentile questions in this section are all written the same way. Each piece of the statement carries a decision, so read it piece by piece before you compute.
Annotate
On: \( P(\bar{X} < k) = 0.90 \quad \Rightarrow \quad k = \mu + z_{0.90}\,\frac{\sigma}{\sqrt{n}} \)
The calculator's invNorm(area, centre, spread) does the last two pieces in one go, which is why the book's solutions are one line long. The deck writes them out, so that you can see which centre and which spread went in; that is where the wrong-distribution errors hide.
\[ z_{0.25} = -0.6745, \quad z_{0.50} = 0, \quad z_{0.75} = 0.6745 \]
\[ z_{0.90} = 1.2816, \quad z_{0.95} = 1.6449 \]
Those five z-values cover every percentile in this section. The quartiles are symmetric about zero, and the median of any normal quantity is simply its centre.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 375 — Example 7.8(b): find k where P(x̄ < k) = 0.90
Worked example
A robbery occurs every two minutes on average, with standard deviation 0.5 minutes; the sample size is 100. Find (a) the median and quartiles of the sample mean, (c) the probability the mean is between 1.75 and 1.85, (d) the value two standard errors above the mean.
Centre and standard error
Why: 0.5 over root 100.
\[ \mu_{\bar{x}} = 2, \quad \sigma_{\bar{x}} = \frac{0.5}{\sqrt{100}} = \frac{0.5}{10} = 0.05 \]
The book says a robbery occurs every two minutes on average, which means the mean time between robberies is 2 minutes. The sample is 100 of those gaps, so the questions are about their mean and their total.
(a) Median
Why: the curve is symmetric, so the median is the centre.
\[ \text{50th percentile} = 2 \]
First quartile
Why: z of minus 0.6745.
\[ 2 + (-0.6745)(0.05) = 1.9663 \approx 1.97 \]
Third quartile
Why: z of plus 0.6745.
\[ 2 + (0.6745)(0.05) = 2.0337 \approx 2.03 \]
Part (c) asks, in the book's words, for the probability that a robbery occurs on the average between 1.75 and 1.85 minutes. On the average signals the mean of the 100 sample times, so it uses the curve with spread 0.05.
(c) Standardise both ends
Why: 1.75 and 1.85.
\[ z_{1} = \frac{1.75 - 2}{0.05} = -5, \quad z_{2} = \frac{1.85 - 2}{0.05} = -3 \]
Area between
Why: the whole band is beyond three standard errors below.
\[ P(-5 < Z < -3) = 0.0013 \]
Figure (svg): A bell curve for the mean time between robberies, centred at 2 with standard error 0.05, with a z row: 1.75 is z minus 5, 1.85 is minus 3, 2.1 is 2. The band from 1.75 to 1.85 is shaded red but is almost invisible against the axis, labelled area 0.0013. A yellow line at 2.1 is labelled two standard errors up.
(d) Two standard errors above
Why: solve the z-score equation for the mean.
\[ \bar{x} = \mu + z\sigma_{\bar{x}} = 2 + 2(0.05) = 2.1 \]
In context: in half of all samples of 100 robberies the mean gap is under 2 minutes, and the middle half of sample means lie between about 1.97 and 2.03 minutes. A sample averaging between 1.75 and 1.85 minutes would be very unusual, about 13 times in 10000.
Check the quartiles by symmetry
Why: they must sit the same distance either side of 2.
\[ 2 - 1.9663 = 0.0337 = 2.0337 - 2 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 378 — Example 7.10(a), (c), (d)
Worked example
Now the sum of the 100 sample times: find its median and quartiles, and the interquartile range. The book writes the sum's standard deviation with the label sigma sub mu x, a misprint for sigma sub sum x; the number, 5, is right.
Centre of the sum
Why: n times μ.
\[ \mu_{\Sigma x} = 100(2) = 200 \]
Spread of the sum
Why: root n times σ.
\[ \sigma_{\Sigma x} = \sqrt{100}(0.5) = 10(0.5) = 5 \]
Median
Why: the centre, by symmetry.
\[ \text{50th percentile} = 200 \]
First quartile
Why: invNorm(0.25, 200, 5).
\[ 200 + (-0.6745)(5) = 196.63 \]
Third quartile
Why: invNorm(0.75, 200, 5).
\[ 200 + (0.6745)(5) = 203.37 \]
The median needed no calculation, and the quartiles sit symmetrically about it, 3.37 minutes either side. Any quartile question on a normal curve reduces to the centre plus or minus 0.6745 spreads.
(e) Interquartile range
Why: third quartile minus first.
\[ \text{IQR} = 203.37 - 196.63 = 6.74 \]
Figure (svg): A bell curve for the mean time between robberies, centred at 2 minutes with standard error 0.05, with the middle 50 percent between 1.966 and 2.034 shaded. A second row under the axis multiplies each tick by 100: 190, 196.63, 200, 203.37, 210, the matching quartiles of the sum.
In context: the middle half of all totals of 100 robbery gaps lie between 196.63 and 203.37 minutes, a span of 6.74 minutes. The IQR of a normal distribution is always 1.349 standard deviations, and 1.349 times 5 is 6.745, which the book's subtraction of rounded values shows as 6.74.
Check against part (a)
Why: each quartile of the sum must be 100 times the quartile of the mean.
\[ 100 \times 1.9663 = 196.63, \quad 100 \times 2.0337 = 203.37 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 378 — Example 7.10(b), (e)
Step zero
Example 7.11 studies the ages of Broadway show attendees: ages range from 14 to 61, with mean 30.9 and standard deviation 9 years. It has six parts, mixing samples of 25, 49 and 65, means and sums. Before any arithmetic, classify each part.
Discussion prompt
For each part, say whether it is about a mean or a sum, give n, and write the centre and spread you would use: (a) the mean age of 25 is under 35; (b) the mean age of 25 is over 50; (c) the sum of 49 ages is at least 1600; (d) the sum of 49 ages is at most 1595; (e) the 95th percentile of the mean of 65; (f) the 90th percentile of the sum of 65.
Setting up all six distributions first catches the two slips this example invites: reusing the spread from one part in another with a different n, and using a mean's spread for a sum. The book itself rounds 9 over root 65 to 1.1 in part (e), which happens not to change the answer.
Is n large enough? The ages run from 14 to 61 around a mean of 30.9, so the population is probably skewed to the right, but samples of 25 or more are usually enough for a moderately skewed population.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 379 — Example 7.11
Worked example
In a sample of 25 attendees, find (a) the probability the mean age is less than 35, and (b) whether a mean age above 50 is likely.
\[ \bar{X} \sim N(30.9,\ 1.8), \quad \sigma_{\bar{x}} = \frac{9}{\sqrt{25}} = 1.8 \]
(a) Standardise 35
Why: 4.1 years above the centre.
\[ z = \frac{35 - 30.9}{1.8} = 2.28 \]
Area to the left
Why: normalcdf(minus E99, 35, 30.9, 1.8).
\[ P(\bar{X} < 35) = 0.9886 \]
Part (b) asks whether a mean above 50 is likely. Before computing, notice that 50 is more than 19 years above 30.9, and the standard error is only 1.8 years. The answer will be no, and the z-score says by how much.
(b) Standardise 50
Why: 19.1 years above the centre.
\[ z = \frac{50 - 30.9}{1.8} = 10.6 \]
Area to the right
Why: more than ten standard errors out.
\[ P(\bar{X} > 50) \approx 0 \]
In context: a group of 25 attendees almost certainly has an average age under 35, and it is practically impossible for their average to be over 50. The book adds the important rider: it is still entirely possible for an individual in the group to be over 50, since ages go up to 61.
That rider is the whole section in one sentence. The mean of 25 is squeezed to within a few years of 30.9, but the people are not; a 55-year-old can be in the audience without dragging the group average anywhere near 50.
Check the scale
Why: the area 0.9886 must match z = 2.28, just under the 99th percentile z of 2.33.
\[ 2.28 < 2.33 \Rightarrow P < 0.99 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 379 — Example 7.11(a)(b)
Worked example
In a sample of 49 attendees, find (c) the probability that the sum of the ages is no less than 1600, and (d) whether the sum is likely to be at most 1595.
Centre of the sum
Why: 49 times 30.9.
\[ \mu_{\Sigma x} = 49(30.9) = 1514.1 \]
Spread of the sum
Why: root 49 is 7.
\[ \sigma_{\Sigma x} = \sqrt{49}(9) = 7(9) = 63 \]
A sum of 49 ages is centred near 1500 years, which sounds strange for ages until you remember it is 49 people added together. The scale check is quick: 1600 divided by 49 is a mean of 32.7, only a little above 30.9.
(c) Standardise 1600
Why: no less than means at least.
\[ z = \frac{1600 - 1514.1}{63} = 1.363 \]
Area to the right
Why: normalcdf(1600, E99, 1514.10, 63).
\[ P(\Sigma X \ge 1600) = 0.0864 \]
(d) Standardise 1595
Why: at most means to the left.
\[ z = \frac{1595 - 1514.1}{63} = 1.284 \]
Area to the left
Why: normalcdf(minus E99, 1595, 1514.10, 63).
\[ P(\Sigma X \le 1595) = 0.9005 \]
Figure (svg): A bell curve for the total age of 49 attendees, centred at 1514.1 years with standard deviation 63, ticks one standard deviation apart with a z row. Everything left of 1595 is shaded blue, labelled at most 1595: 0.9005. The tail right of 1600 is shaded red, labelled at least 1600: 0.0864. The two cut lines are only five years apart.
In context: there is about an 8.6 percent chance that 49 attendees have a combined age of at least 1600 years, and about a 90 percent chance that their combined age is at most 1595 years. So yes, a total of at most 1595 is likely.
Check that the two parts nearly complement
Why: together with the sliver between 1595 and 1600 they must make exactly 1.
\[ P(1595 < \Sigma X < 1600) = 0.0131 \]
\[ 0.9005 + 0.0131 + 0.0864 = 1.0000 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 379 — Example 7.11(c)(d)
Worked example
For samples of 65 attendees, find (e) the 95th percentile of the sample mean age and (f) the 90th percentile of the sum of the ages, and interpret each.
(e) Standard error
Why: the book rounds this to 1.1.
\[ \sigma_{\bar{x}} = \frac{9}{\sqrt{65}} = \frac{9}{8.062} = 1.1163 \]
Walk 1.645 standard errors up
Why: invNorm(0.95, 30.9, 1.1163).
\[ k = 30.9 + 1.6449(1.1163) = 32.74 \approx 32.7 \]
With the book's rounded 1.1 the cut is 32.71, so the rounding does not change the printed 32.7. Now part (f), which is a sum of the same 65 ages, so both the centre and the spread change scale.
(f) Centre and spread of the sum
Why: 65 times 30.9, and root 65 times 9.
\[ \mu_{\Sigma x} = 2008.5, \quad \sigma_{\Sigma x} = 8.062(9) = 72.56 \]
Walk 1.2816 standard deviations up
Why: invNorm(0.90, 2008.5, 72.56).
\[ k = 2008.5 + 1.2816(72.56) = 2101.5 \]
Now the interpretation, and here the book slips. It says 95 percent of the attendees in the sample of 65 are younger than 32.7 years on average. But a percentile of the mean describes the means of all possible samples, not the people in one sample. The correct sentence: 95 percent of all samples of 65 attendees have a mean age below 32.7 years.
Likewise for (f): 90 percent of all samples of 65 attendees have a combined age below 2101.5 years. It is not that 90 percent of the attendees have some sum of ages; a single attendee has an age, not a sum.
Check (f) against the mean's scale
Why: divide by 65: the 90th percentile of the mean should come out just below the 95th, 32.74.
\[ \frac{2101.5}{65} = 32.33 = 30.9 + 1.2816(1.1163) \quad \checkmark \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 379 — Example 7.11(e)(f)
Concept
The book asks you to interpret every percentile in a complete sentence, and its own sentences show how easy it is to get the subject wrong. A reliable sentence has four slots, and each one is a decision you have already made.
| slot | what goes in it | Example 7.8(b) |
|---|---|---|
| how much | the percentile as a percentage | 90 percent |
| of what | all samples of size n, and which statistic | of all means of 75 stress scores |
| which side | at most, or below, for a percentile | are at most |
| the cut, with units | k in the problem's units | 3.2 points |
Put together: 90 percent of all means of 75 stress scores are at most 3.2. The second slot is the one that fails. It must name the samples and the statistic, because a percentile of a mean describes the collection of means you would get from many samples, not the people in any one of them.
A probability gets the same treatment. For Example 7.11(a): there is a 0.9886 probability that the mean age of a random sample of 25 attendees is less than 35 years. Swap the subject for the attendees themselves and the sentence becomes false, since plenty of individual attendees are over 35.
One more habit: the book often adds the complementary half, and 10 percent are at least 3.2. It costs one clause and shows you know which way the area runs.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, pp. 375-379 — the interpretations in Examples 7.8, 7.9 and 7.11
Trap
Example 7.9 found the 95th percentile of the sample mean excess time for 80 customers. A first interpretation, like the book's own wording in Example 7.11(e), turns it into a statement about customers:
\[ k = 26.0 \text{ min} \]
Wrong: 95 percent of customers go less than 26 minutes over.
The percentile belongs to the distribution of sample means. It describes samples of 80, not the customers inside them.
Right: 95 percent of samples of 80 customers have a mean excess time under 26 minutes.
For individual customers the figure is very different. One customer's excess time is exponential, so the share going more than 26 minutes over is:
\[ P(X > 26) = e^{-26/22} = 0.307 \]
So about 31 percent of customers exceed 26 minutes, not 5 percent. The fix is a habit: name the quantity in every interpretation sentence, the mean of 80, or the total of 65, or one customer.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, pp. 378-379 — Example 7.9 and Example 7.11(e)
Matching
Every number below came out of this section. Match it with the sentence that interprets it correctly. Watch for the scale, and for whether the sentence is about one value, a mean, or a total.
Match the pairs
Why: The scale sorts most of these at once. A stress-score mean lives between 1 and 5, so 3.2 goes with the means; a total of 75 scores is in the hundreds, so 237.8 goes with the totals. The two probabilities differ in subject: 0.4029 is about one customer, 0.9886 about a group's average. The IQR is a length on the total's scale.
If you swapped 3.2 and 237.8, go back to the scale check: 237.8 is not a possible mean of stress scores. If you paired 0.4029 with a mean, remember that the mean of 80 customers exceeds 20 minutes with probability 0.7919, not 0.4029.
The habit these pairs train is the one the book asks for: every number gets a sentence, and every sentence names what the number is about.
Section
Part 6
Concept
The section ends with a historical note. Before calculators, binomial probabilities with large n were hard to compute, and the central limit theorem gave a shortcut: approximate the binomial by a normal curve.
Why does the theorem apply? A binomial count X is the number of successes in n independent trials. Score each trial 1 for a success and 0 for a failure; then X is the sum of those n zeros and ones. A sum of many independent values is exactly what the central limit theorem for sums is about.
\[ \mu = np, \qquad \sigma = \sqrt{npq}, \qquad q = 1 - p \]
Figure (svg): Bars for the binomial probabilities of 130 to 188 people in favour out of 300 with p equal to 0.53, rising to a peak of about 0.046 at 159 and falling symmetrically. A yellow normal curve with mean 159 and standard deviation 8.64 runs across the tops of the bars almost exactly.
The condition is that the binomial's shape is close enough to normal: both np and nq must be greater than 5, and the book notes the approximation is better when both are at least 10. Here n is 300 and p is 0.53, so np is 159 and nq is 141, far above either threshold.
\[ np = 300(0.53) = 159, \quad nq = 300(0.47) = 141, \quad \sigma = \sqrt{74.73} = 8.6447 \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, pp. 379-380 — Historical note: normal approximation to the binomial
Concept
A binomial count takes only whole-number values, but the normal curve is continuous. To use one for the other, think of each whole number as a bar one unit wide, centred on that number.
Figure (svg): A close-up of binomial bars from 144 to 156, each one unit wide and centred on its whole number. Bars for 150 and above are yellow, bars below 150 are blue. A dashed line at 149.5 marks the left edge of the bar for 150, and the area under the yellow normal curve to the right of 149.5 is shaded.
The bar for 150 runs from 149.5 to 150.5. So at least 150 means all the bars from 150 upward, and their area under the curve starts at 149.5. Moving the cut by 0.5 to cover the whole bar is the continuity correction.
\[ P(X \ge 150) \approx P(Y \ge 149.5) \]
The rule is always to ask whether the boundary value is included. If it is, as in at least or at most, extend the cut half a unit outward to include its whole bar. If it is excluded, as in more than or fewer than, move half a unit inward to leave its bar out. Exactly one value becomes the band from half below to half above.
\[ P(X = 175) \approx P(174.5 < Y < 175.5) \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 380 — the continuity correction factor
Notation
The whole method fits in one line, and each part of it has a condition or a reason attached. Read it slowly before the worked examples use it five times.
Annotate
On: \( P(X \ge a) \approx P(Y \ge a - 0.5), \quad Y \sim N\left(np,\ \sqrt{npq}\right) \)
Notice that nothing here is new except the half unit. The centre and spread are the binomial's own, from Chapter 4, and the normal curve is used exactly as in Chapter 6. The central limit theorem is what licenses it, because a count of successes is a sum of zeros and ones.
\[ np = 159 > 5, \quad nq = 141 > 5 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 380 — μ = np, σ = √(npq), and the continuity correction factor
Sorting
Each binomial event below is about the charter-school survey, with X the number in favour. Decide whether its normal cut moves half a unit to include the boundary value's bar or to exclude it. Say whether the boundary is included first.
Sort into buckets
Sort each event by what the continuity correction does to its boundary.
A quick way to get it right every time: list the first whole number that is in the event, then take the edge of its bar. For more than 155, the first value in is 156, and its bar starts at 155.5.
Words like no more than and over are there to test that you translate to an inequality first. No more than 170 means at most 170, so it includes 170.
Worked example
In a school district, 53 percent favour a charter school. A simple random sample of 300 is surveyed. Find the probability that (a) at least 150 favour it, (b) at most 160 favour it, using the normal approximation.
\[ X \sim B(300,\ 0.53) \approx Y \sim N(159,\ 8.6447) \]
Check the conditions first: np is 159 and nq is 141, both far above 5, so the normal curve fits the binomial well, as the bar picture showed. Its centre is 159 and its spread 8.6447.
(a) Continuity correction
Why: 150 is included, so start at its bar's left edge.
\[ P(X \ge 150) \approx P(Y \ge 149.5) \]
Standardise
Why: 9.5 below 159.
\[ z = \frac{149.5 - 159}{8.6447} = -1.099 \]
Area to the right
Why: normalcdf(149.5, 10^99, 159, 8.6447).
\[ P(Z \ge -1.099) = 0.8641 \]
The z-score is negative because 149.5 is below the centre of 159, so the area to its right is more than a half. Now part (b), where the boundary is again included, but on the other side.
(b) Continuity correction
Why: 160 is included, so end at its bar's right edge.
\[ P(X \le 160) \approx P(Y \le 160.5) \]
Standardise
Why: 1.5 above 159.
\[ z = \frac{160.5 - 159}{8.6447} = 0.174 \]
Area to the left
Why: normalcdf(0, 160.5, 159, 8.6447).
\[ P(Z \le 0.174) = 0.5689 \]
In context: there is about an 86 percent chance that at least 150 of the 300 favour a charter school, and about a 57 percent chance that at most 160 do. Without the correction, part (a) would start at 150 and give 0.8511, missing half of the 150 bar.
Check against the exact binomial
Why: node sums the binomial probabilities directly, as the book does with binomialcdf.
\[ 1 - \text{binomcdf}(300, 0.53, 149) = 0.8641, \quad \text{binomcdf}(300, 0.53, 160) = 0.5684 \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, pp. 380-381 — Example 7.12(a)(b)
Worked example
Same survey, X the number of the 300 in favour. Find the probability that (c) more than 155 favour, (d) fewer than 147 favour, (e) exactly 175 favour.
(c) Excluded boundary
Why: the first value in is 156, whose bar starts at 155.5.
\[ P(X > 155) \approx P(Y > 155.5) \]
Standardise and take the right area
Why: 3.5 below 159.
\[ z = \frac{155.5 - 159}{8.6447} = -0.405, \quad P = 0.6572 \]
More than 155 is a strict inequality, so 155 itself is out and the cut moves up by half a unit rather than down. Fewer than 147 is strict too, so its cut moves down from 147 to 146.5.
(d) Excluded boundary
Why: the last value in is 146, whose bar ends at 146.5.
\[ P(X < 147) \approx P(Y < 146.5) \]
Standardise and take the left area
Why: 12.5 below 159.
\[ z = \frac{146.5 - 159}{8.6447} = -1.446, \quad P = 0.0741 \]
(e) One value is one bar
Why: from 174.5 to 175.5.
\[ P(X = 175) \approx P(174.5 < Y < 175.5) \]
Standardise both edges and subtract
Why: a thin slice far up the right side.
\[ z = 1.793 \text{ to } 1.909, \quad P = 0.0083 \]
Part (e) shows why the correction is not optional. Without it, exactly 175 would be the area over a single point, which for a continuous curve is zero. The half unit either side turns the point back into the bar it stands for.
Check against the exact binomial
Why: the book lists these values; node agrees.
\[ 0.6576, \quad 0.0742, \quad 0.0083 \quad \text{vs normal } 0.6572, \ 0.0741, \ 0.0083 \quad \checkmark \]
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, pp. 380-381 — Example 7.12(c)-(e)
Trap
Example 7.12(a) asks for at least 150 of 300. A first attempt standardises 150 itself:
\[ z = \frac{150 - 159}{8.6447} = -1.041 \]
\[ P(Z \ge -1.041) = 0.8511 \]
Wrong by 0.013: half of the bar for 150 is missing.
The count 150 is a bar from 149.5 to 150.5, and at least 150 includes all of it. Start at its left edge.
\[ z = \frac{149.5 - 159}{8.6447} = -1.099 \]
\[ P(Z \ge -1.099) = 0.8641 \]
That matches the exact binomial, 0.8641, to four places. The error without the correction looks small here because n is large, but for exact values like part (e) it is total: the uncorrected answer is zero.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 380 — Example 7.12(a)
Section
Part 7
Pattern
Every problem in this section ran through the same six steps. The arithmetic in step five is Chapter 6 again; steps one to four and step six are where this section lives.
Figure (svg): Six boxes in a row joined by arrows: define X as one value in words; find mu and sigma of the population; decide individual, mean or sum, highlighted in yellow and labelled the deciding step; check that n is large or X is normal; compute the z-score, area or cut; write the sentence in context.
\[ \text{individual: } X \quad \big| \quad \text{mean: } N\left(\mu, \tfrac{\sigma}{\sqrt{n}}\right) \quad \big| \quad \text{sum: } N(n\mu, \sqrt{n}\,\sigma) \]
The step people skip is the third. Every error in this lesson that produced a confident wrong number, the normal curve for one customer, the total's spread used for a mean, the door designed for the average man, came from computing before deciding.
Two built-in checks catch most errors. The scale check: a mean must be in the range of the data, and a total must be n times that range. The consistency check: any answer about a total, divided by n, must match the same answer about the mean.
Check
An unknown population has mean 50 and standard deviation 12, and you take a sample of 36. Decide which distribution the question needs before you compute.
Check your understanding
A population of unknown shape has mean 50 and standard deviation 12. For a random sample of 36, what is the probability that the sample mean is greater than 52?
Answer: A
Why: The question is about a mean, and 36 is large, so the sample mean is approximately normal with mean 50 and standard error 12 over 6, which is 2. The value 52 is one standard error above 50, and the area to the right of z equal to 1 is 0.1587.
\[ \sigma_{\bar{x}} = \frac{12}{\sqrt{36}} = 2, \quad z = \frac{52 - 50}{2} = 1, \quad P(Z > 1) = 0.1587 \]
In context: the probability that a random sample of 36 from this population has a mean above 52 is about 0.16, roughly one sample in six.
Every wrong option is a slip from this lesson: the individual's spread, the square root forgotten, and the wrong tail. The shape of the population does not matter here because n is 36; it would matter for a question about one value, which could not be answered at all without knowing the shape.
Check
Try It 7.12: 46 percent of a city favour the incumbent, Dawn Morgan, and a simple random sample of 500 is taken. Use the continuity correction.
Check your understanding
Using the normal approximation with the continuity correction, what is the probability that at least 250 of the 500 favour Dawn Morgan?
Answer: A
Why: The mean is 500 times 0.46, which is 230, and the standard deviation is the square root of 500 times 0.46 times 0.54, which is 11.14. At least 250 includes the bar for 250, so the cut is 249.5, which is 1.75 standard deviations above 230, and the area to the right is 0.0401.
\[ \mu = 230, \quad \sigma = \sqrt{124.2} = 11.14, \quad z = \frac{249.5 - 230}{11.14} = 1.75 \]
The exact binomial probability, summed in node, is 0.0402, so the approximation is excellent. In context: if 46 percent of the city favours her, a sample of 500 would show 250 or more supporters only about 4 percent of the time.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, p. 381 — Try It 7.12
Check
The last check is about the sentence, not the number. Read every option for what it claims to describe.
Check your understanding
For samples of 80 customers who exceed their contract minutes, the 95th percentile of the sample mean excess time is 26.0 minutes. Which interpretation is correct?
Answer: A
Why: The percentile belongs to the distribution of the sample mean, so it describes the means of samples of 80. Ninety-five percent of those means are below 26.0 minutes and five percent are above.
Each wrong option swaps the subject of the sentence: individual customers, the fixed population mean, or the members of one sample. Only the sample means have the normal distribution the 26.0 came from.
It is worth being precise about the third option, because it is the error that later chapters on confidence intervals are built to prevent. The population mean is a fixed number; it is the sample means that vary.
A reliable template: so many percent of all samples of size n have a mean below k, in the units of the problem.
Explain it to yourself
The book's note says: if you are asked for the probability of an individual value, do not use the central limit theorem. Explaining why is the best test of whether this section has landed.
Discussion prompt
In a few sentences, explain why the central limit theorem cannot be used for one customer's excess minutes in Example 7.9, but can be used for the mean of 80 customers. Use the two probabilities for 20 minutes, 0.4029 and 0.7919, in your explanation.
A strong answer names three things: what the theorem is about (means and sums of many values), why a single value is not covered (nothing is averaged), and what shape the individual keeps (the population's own). The numbers then illustrate the argument rather than replace it.
If your explanation said the mean is more accurate, push it one step further: more accurate about what? About the population mean, 22. A single customer is not trying to estimate anything; she is one draw from a skewed distribution.
Exit ticket
A help desk's waiting times are exponential with mean 8 minutes, so the standard deviation is also 8. A manager looks at samples of 64 callers.
Discussion prompt
Find (a) the probability that one caller waits more than 10 minutes, (b) the probability that the mean wait of 64 callers is more than 10 minutes, and (c) the 90th percentile of the total wait of 64 callers. Say which distribution each part uses, and interpret (c) in a sentence.
The marks are in the decisions. Part (a) earns nothing with a normal curve, because one caller is exponential. Parts (b) and (c) use the two central limit theorem distributions, and the check that ties them is 594.0 divided by 64, which is 9.28, exactly 8 plus 1.2816 standard errors.
If the manager's target is that callers should not wait more than 10 minutes, notice which number protects callers: the 29 percent from part (a), not the 2 percent from part (b).
Recap
One population supports three distributions, and the first job in every problem is to pick the right one. The deciding word is in the question: one value, a mean, or a total.
| the question is about | use | centre | spread |
|---|---|---|---|
| one value (one, a randomly selected) | X's own distribution | μ | σ, shape of the population |
| a mean (mean, average) | CLT: x̄ ~ N(μ, σ/√n) | μ | σ/√n, shrinks as n grows |
| a sum (total, sum) | CLT: ΣX ~ N(nμ, √n·σ) | nμ | √n·σ, grows as n grows |
| a binomial count, np and nq > 5 | N(np, √(npq)), ± 0.5 | np | √(npq) |
The law of large numbers says sample means close in on μ as n grows, and the central limit theorem shows how fast: the standard error shrinks like one over root n. That is why the mean of 80 skewed excess times is a narrow bell while one customer stays exponential, and why their answers for 20 minutes, 0.79 and 0.40, are both right.
Percentiles of totals are n times percentiles of means, which gives a free check. And every answer ends in a sentence that names its subject: all samples of 80, the total of 75, one customer. Introductory Business Statistics works through the same material with business contexts, if you want more practice.
OpenStax Introductory Statistics 2e, §7.3 Using the Central Limit Theorem §7.3, pp. 373-381 — the whole section
OpenStax Introductory Business Statistics 2e, §7.2 Using the Central Limit Theorem §7.2 Using the Central Limit Theorem — the same ideas in business contexts
Want this taught 1-on-1? Alexander tutors Statistics — $55/session, free consultation.